Deep-Reinforcement-Learning-Based Object Transportation Using Task Space Decomposition.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Traditional deep reinforcement learning (DRL) for object transportation is limited to specific, small environments.
• DRL methods often struggle with convergence in complex or large-scale settings due to environmental dependency.
• Existing approaches require extensive retraining for new or varied environments.

Purpose of the Study:

• To propose a novel DRL-based object transportation method overcoming limitations of prior approaches.
• To enhance the adaptability and scalability of DRL in robotic manipulation tasks.
• To enable robots to transport objects in large and complex environments without re-learning.

Main Methods:

• The study employs task space decomposition (TSD) to break down complex transportation tasks into simpler sub-tasks.
• Robots first learn object transportation in a standard learning environment (SLE) with simple structures.
• The task space is decomposed into sub-task spaces with defined sub-goals, which the robot sequentially achieves.

Main Results:

• The proposed method demonstrated successful object transportation in diverse and complex simulated environments, including corridors, polygons, and mazes.
• The approach allows for seamless extension to new, large-scale environments without additional training or re-learning.
• Simulations verified the method's effectiveness and robustness across varied scenarios.

Conclusions:

• The novel DRL with TSD method significantly improves the scalability and applicability of robotic object transportation.
• This approach addresses the limitations of conventional DRL methods in complex and dynamic environments.
• The method offers a robust solution for real-world robotic manipulation tasks requiring adaptability.